mimic-generate

Two phases:

• Phase A — brief synthesis (you produce the AdBrief in your response, no API call).
• Phase B — image render (Bash + curl to the chosen provider, download to disk).

Each phase persists the Generation JSON to disk so the dashboard reflects in-progress work.

Inputs

• brief — slug (Wiki/Mimic/Briefs/{brief}.md must exist)
• template — slug (Wiki/Mimic/Templates/{template}.md must exist)
• product, persona, angle, offer — free-text
• size — 1:1 | 3:4 | 16:9
• framework — standard | rmbc2
• provider — kie | fal | openai | google | openrouter
• model — provider-specific identifier

Step 0 — Generate the ID

id = "mim_" + YYYY-MM-DD + "_" + 6 random alphanumeric chars

Use Bash to mint randomness if needed:

echo "mim_$(date -u +%Y-%m-%d)_$(LC_ALL=C tr -dc 'a-z0-9' </dev/urandom | head -c6)"

Step 1 — Read brief + template

Read Wiki/Mimic/Briefs/{brief}.md     → parse frontmatter → BrandDNABrief
Read Wiki/Mimic/Templates/{template}.md → parse frontmatter → AdTemplate

Phase A — Brief synthesis

Generate the AdBrief JSON in your response. The brief is yours to author — you have all context. Schema:

{
  "headline_options": ["3-5 alternative headlines, max ~10 words each"],
  "subheadline_options": ["3-5 subheads, ~12-15 words each"],
  "badges": ["short badge text strings (e.g. 'AS LOW AS $24.95/mo')"],
  "cta": "single CTA text (e.g. 'GET QUOTE')",
  "layout_instructions": "where each piece goes — references the template's locked_structural_elements",
  "image_direction": "designer-facing description of the hero visual",
  "image_generation_prompt": "the actual prompt that goes to the image provider — should fill in the template's image_prompt_template placeholders with brand DNA + product values",
  "designer_notes": "anything else a designer would need (typography weight, color usage, photo treatment)",
  "compliance_notes": "what claims this ad makes + which disclaimers apply (cite from brief.claims_guardrails)",
  "rmbc2": null   // only when framework === "rmbc2", populate per RMBC2Layer schema below
}

RMBC II layer (when framework === "rmbc2")

Stefan Georgi's Research → Mechanism → Brief → Copy. Add this object on the AdBrief:

{
  "research_synthesis": "2-4 sentences distilling Brand DNA + persona + angle inputs",
  "awareness_level": "unaware | problem-aware | solution-aware | product-aware | most-aware",
  "awareness_rationale": "why you placed the persona at this level",
  "ump": {
    "name": "Memorable name for the Unique Mechanism of Problem (e.g. 'Collagen Architecture Breakdown at the Dermal-Epidermal Junction')",
    "explanation": "what the actual underlying problem is, mechanically"
  },
  "ums": {
    "name": "Memorable name for the Unique Mechanism of Solution (e.g. 'The Structural Renewal System')",
    "explanation": "what your solution does mechanically",
    "why_yours_unique": "why this version vs competitors"
  },
  "big_idea": "the lead angle in one line",
  "hook_variations": [
    { "type": "problem-agitation | curiosity | transformation | status | scarcity | social-proof | contrast | founder-story | callout | listicle | before-after | fear-removal", "hook": "..." }
    // 4-6 hooks across hook types so Adam can A/B test
  ],
  "proof_layer": "why-believe-this anchor — testimonial / stat / certification",
  "primary_objection": "the #1 objection this persona will raise",
  "objection_demolish": "how the ad pre-handles the objection visually + verbally"
}

Synthesis rules

• Honor template.template_logic.locked_structural_elements — the new ad must keep the same structural layout (same hook position, same badge structure, same CTA shape).
• Honor template.template_logic.editable_fields — those are what you swap.
• Strict on brief.banned_phrases — none of those words appear anywhere in the AdBrief.
• Strict on brief.claims_guardrails — never use a cannot_claim phrasing. If you use a requires_disclaimer claim, add the disclaimer to compliance_notes.
• approved_phrases preferred — incorporate where natural.

Persist Phase A

Write the Generation JSON now (status = drafted, image = null):

mkdir -p /Users/doorsy/Desktop/supernova/supernova-ui/data/mimic/generations
# Write JSON via Write tool to:
# /Users/doorsy/Desktop/supernova/supernova-ui/data/mimic/generations/{id}.json

Generation JSON shape:

{
  "id": "mim_2026-05-08_xxxxxx",
  "created": "2026-05-08T16:42:00.000Z",
  "framework": "standard | rmbc2",
  "brand_brief_slug": "{brief}",
  "template_slug": "{template}",
  "inputs": { "product": "...", "persona": "...", "angle": "...", "offer": "..." },
  "size": "1:1 | 3:4 | 16:9",
  "ad_brief": { /* AdBrief object */ },
  "image": null,
  "status": "drafted"
}

Surface to Adam: "Brief drafted, dashboard shows it now. Rendering image…"

Phase B — Image render

Provider API specs

Kie (async task pattern)

Create task: POST https://api.kie.ai/api/v1/jobs/createTask

curl -s -X POST https://api.kie.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $KIE_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$(cat <<EOF
{
  "model": "{model}",
  "input": {
    "prompt": {prompt},
    "aspect_ratio": "{1:1|3:4|16:9}",
    "resolution": "{1K if 1:1 else 2K}",
    "output_format": "png"
  }
}
EOF
)"

Response: {"code":200,"data":{"taskId":"..."}}

Poll: GET https://api.kie.ai/api/v1/jobs/recordInfo?taskId={id} every 2.5s until data.state === "success" or "fail". On success: data.resultJson is a JSON string with { resultUrls: ["..."] }.

Kie models: nano-banana-2, nano-banana-pro, gpt-image-2-text-to-image.

Fal (synchronous)

POST https://fal.run/{model} with Authorization: Key $FAL_KEY. Body shape branches by model family:

• fal-ai/nano-banana-2 / fal-ai/nano-banana-pro: { prompt, aspect_ratio: "1:1|16:9|3:4", resolution: "1K|2K", output_format: "png", safety_tolerance: "4", num_images: 1 }
• fal-ai/gpt-image-2: { prompt, image_size: "square_hd|portrait_4_3|landscape_16_9", quality: "high", num_images: 1, output_format: "png" }
• fal-ai/flux/dev: { prompt, image_size, num_images: 1, enable_safety_checker: true }

Response: { "images": [{ "url": "..." }] }.

OpenAI direct

POST https://api.openai.com/v1/images/generations with Authorization: Bearer $OPENAI_API_KEY.

{ "model": "gpt-image-2", "prompt": "...", "n": 1, "size": "1024x1024 | 1024x1536 | 1536x1024", "quality": "high" }

Response: { "data": [{ "url": "..." }] } (or { "b64_json": "..." }).

Size mapping: 1:1 → 1024x1024, 3:4 → 1024x1536, 16:9 → 1536x1024.

OpenAI may require Org Verification before allowing gpt-image-2.

Google GenAI (Imagen 4)

POST https://generativelanguage.googleapis.com/v1beta/models/{model}:generateImage?key=$GOOGLE_GENAI_API_KEY. Refer to Google's docs if used — rare path.

OpenRouter

POST https://openrouter.ai/api/v1/images/generations with Authorization: Bearer $OPENROUTER_API_KEY and HTTP-Referer + X-Title headers. Model openai/gpt-image-2. Same body as OpenAI direct.

Download the image

Whichever provider returned a URL:

mkdir -p /Users/doorsy/Desktop/supernova/supernova-ui/data/mimic/images
curl -s -L -o /Users/doorsy/Desktop/supernova/supernova-ui/data/mimic/images/{id}.{ext} "{image_url}"

If the provider returned base64 (b64_json): decode via base64 -d into the same path.

{ext} — png if output_format: png (most providers default). Match the actual content-type if returned.

Update the Generation JSON

Re-Write the same {id}.json with status complete:

{
  ...all original fields...,
  "image": {
    "provider": "{provider}",
    "model": "{model}",
    "image_path": "data/mimic/images/{id}.{ext}",
    "image_url": "{remote URL if available}"
  },
  "status": "complete"
}

If render failed:

{
  ...all original fields...,
  "image": {
    "provider": "{provider}",
    "model": "{model}",
    "error": "{error message}"
  },
  "status": "failed"
}

Never delete the Generation JSON — failed-state generations are visible in the gallery and can be retried.

Step 5 — Log

Invoke Skills/log-operation.md with operation mimic-generate, generation_id, brief_slug, template_slug, framework, provider, model, status, image_path (if complete), error (if failed).

Report to Adam

✓ Generated mim_2026-05-08_xxxxxx
   brief: {brief} · template: {template} · {framework} · {size}
   provider: {provider} ({model})
   status: complete | failed
   image: data/mimic/images/{id}.{ext} (or error: ...)
   Dashboard: http://localhost:3001/mimic/generations/{id}

If compliance_notes flagged anything, surface that in the report so Adam knows what disclaimers must be added in design.